PRESELECTION OF NEUROSTIMULATION WAVEFORMS FOR

VISUAL PROSTHESES USING GENETIC ALGORITHMS

Samuel Romero, Alberto Guill

en, Crist

obal J. Carmona

Department of Computer Science, University of Jaen, Spain

C. Morillas, F. Pelayo, H. Pomares

Department of Computer Technology and Architecture, University of Granada, Spain

Keywords:

Visual prostheses, Electrical neurostimulation, Bioelectric waveform, Genetic algorithms.

Abstract:

Among the variety of approaches for developing therapies for the blind, electrical neurostimulation of the

visual pathways seems to be a promising choice. Delivering bi-phasic bioelectric pulses to the nerves implies

the selection of values for a number of parameters within a wide range. This needs to be done for every im-

planted electrode, and for every patient. Nowadays, electrode arrays can include up to one hundred channels,

and we expect to raise to thousands of them in a near future. This unavoidable task becomes extremely time-

consuming both for the researcher and for the patient. Therefore, in order to reduce the number of tests to be

carried out in vivo, we propose the use of multi-objective genetic algorithms that can provide a limited set of

candidate waveforms to be tried.

1 INTRODUCTION

According to estimations by the World Health Orga-

nization (WHO, 2005), 45 million people are legally

blind, and more than 135 million persons suffer from

low vision. The number of patients losing their visual

function is expected to rise in the next years, as many

of the illnesses causing blindness are related to aging.

Whenever a person gets blind, severe limitations

in a variety of aspects of his or her life arise. The

patient ﬁnds a number of handicaps in daily activity,

labor, autonomy, social interaction, etc. Additionally,

blindness imposes high costs to public health admin-

istrations.

The term blindness has a functional meaning.

However, a wide variety of traumas and illnesses can

be behind this condition: age-related macular degen-

eration (AMD), diabetic retinopathy, retinitis pigmen-

tosa (RP), cataracts, accidents, etc.

The plethora of resources under research for cur-

ing blindness is even more varied than the illnesses

behind it. Research groups are trying to ﬁnd treat-

ments for visual deprivation from different perspec-

tives: pharmacology, surgery, gene therapy, or even

electrical neurostimulation.

This paper presents an approach to obtain a re-

duced set of stimulation parameters in order to limit

the range of possible combinations of possible values

for electrical neurostimulation. The optimization of

the parameters is done by means of a multiobjective

genetic algorithm which is able to ﬁnd a balance be-

tween the different solutions.

2 VISUAL NEUROPROSTHESES

FOR BLINDNESS

This last bioengineering therapeutic research line tries

to restore a rudimentary, but functional form of visual

perception, by artiﬁcially stimulating part of the valid

remaining neural tissue in the visual pathway.

In this line, several research projects for creat-

ing a visual neuroprosthesis for the blind are cur-

rently under development. Thus, we can mention reti-

nal implants (Humayun, 2003), optic nerve implants

(Veraart, 1998), and cortical implants (Troyk, 2003;

Fern

andez et al., 2005a). All of them try to elicit vi-

sual perceptions by electrically inducing the activa-

tion of nerve cells in the visual pathway.

191

Romero S., Guillén A., J. Carmona C., Morillas C., Pelayo F. and Pomares H. (2010).

PRESELECTION OF NEUROSTIMULATION WAVEFORMS FOR VISUAL PROSTHESES USING GENETIC ALGORITHMS.

In Proceedings of the Third International Conference on Biomedical Electronics and Devices, pages 191-194

DOI: 10.5220/0002744301910194

 SciTePress

Some encouraging results from primitive attempts

can be found in the literature, demonstrating the proof

of concept for visual neuroprosthetic implants (Do-

belle, 2000; Schmidt et al., 1996). In these cases, pa-

tients related the perception of visual sensations when

an electrical pulse was delivered to the implant.

However, a long list of issues needs to be ad-

dressed in order to have a fully usable implant for

the blind: biocompatibility, information processing,

portability, and reliability (Normann et al., 2009).

One of the challenges in these neurostimulation sys-

tems is ﬁnding adequate values for a number of pa-

rameters of the stimulation waveform.

We present in this paper an approach to ﬁnd a

set of candidate waveforms using a multiobjective ge-

netic algorithm. In this way it is possible to ﬁnd a

balance between the different values to be optimized.

Then, the number of trials with patients in clin-

ical research can be signiﬁcantly reduced, as only a

reduced set of speciﬁc neurostimulation waveforms

need to be tested in vivo.

3 ELECTRICAL

NEURO-STIMULATION

WAVEFORMS

In order to activate a neuron in a biological tissue, an

action potential can be generated by putting an elec-

trode in the neighborhood of the cell, and delivering

electrical charges in the medium (charge injection), or

using the same electrode to create an electrical ﬁeld

that elicits the re-distribution of electrical charges in

the surroundings. The effect, under adequate condi-

tions, is the generation of a pulse along the axon of

the neuron, known as action potential. This pulse can

be transmitted to the rest of neurons in the visual path-

way, activating a percept in the visual ﬁeld of the pa-

tient.

Traditionally, neurostimulation pulses are bi-

phasic square waveforms. A negative current at the

beginning induces the effect of the action poten-

tial (”cathodic-ﬁrst stimulation”), while the positive

counter-phase restores electrochemically the neigh-

borhood of the active tip of the electrode. This way,

the amount of charge after the pulse is delivered re-

mains balanced.

Usually, a train of biphasic pulses is required to

elicit a visual percept or ”phosphene” (a perception

similar to a star in the sky).

The set of values for getting a patient seeing a

phosphene is unknown, and needs to be determined

experimentally. Moreover, these values can be differ-

ent for every electrode in an implant, and for every

implanted patient. This means, that testing all possi-

ble combinations of waveform parameter values, for

every electrode in an implant of 100 or even more

electrodes can take an excessive amount of time.

The set of parameters of the stimulation waveform

are the following:

• Current: amplitude of the negative and positive

phases).

• PD: pulse duration (pulse width in the positive or

negative phase).

• IPHI: inter-phase interval (time between negative

an positive phases).

• IPI: inter-pulse interval (time between two con-

secutive biphasic pulses).

• NPulses: number of pulses in a train.

Time parameters are usually expressed in mi-

croseconds, and current is expressed in microam-

peres.

Our objective is to minimize the amount of charge

injected in the neural tissue (to avoid undesirable

harmful effects), which should be over an unknown

threshold in order to elicit a phosphene. Additionally,

we want to maximize the duration of the phosphene

(”onset”), and the brightness of the percept.

In order to enhance the design of a visual pros-

thesis (Fern

andez et al., 2005b), we have developed

a simulator of visual prostheses, based on limited re-

sults of real human implants, from which the rules for

generating the outputs have been extracted.

4 GENETIC ALGORITHMS

APPLIED TO

NEUROSTIMULATION

OPTIMIZATION

Genetic algorithms (GAs) are general purpose search

algorithms which use the principles inspired by nat-

ural genetics to evolve problem solutions (Golberg,

1989). These algorithms have been succesfully ap-

plied to a wide range of real-world problems (Guillen

et al., 2009; Casillas and Carse, 2009; Cord

on et al.,

2007; Guill

en et al., 2006; Isibuchi, 2007) and are

well known in the computer science ﬁeld.

A GA deﬁnes an initial population of individu-

als, each individual encodes a solution to the prob-

lem. The algorithm iterates until a stop criterium is

reached, the sequence of steps that is followed on each

iteration is:

1. evaluate each individual

BIODEVICES 2010 - International Conference on Biomedical Electronics and Devices

192

2. select individuals to be crossed

3. apply the crossover operators to generate the off-

springs

4. select which offsprings and ancestors will form

the next population

5. apply mutation

The idea underneath these algorithms is that if two

good solutions are combined together, the resulting

solutions could be better. The literature regarding

GAs presents a large number of papers performing

studies to the different parameters and proposing new

crossover, mutation, selection policies, etc. Among

this vast variety of GAs, there is a special kind known

as MultiObjective Genetic Algorithms (MOGA). The

main feature of these is that they do not consider a

unique value to determine the quality of a solution (in-

dividual) but use a vector of values. Thus, for some

individuals, it is not possible to say that one individ-

ual is better than the other. A Pareto front is a set of

solutions where each solution is not better than the

others.

In a formal way, a multi-objective op-

timization problem can be deﬁned in

the following way:min/max

−→

y = f (

−→

x ) =

(

−→

x ), f

(

−→

x ),· · · , f

(

−→

x ) where

−→

x = (x

,· · · ,x

)

is a solution and

−→

y = (y

,· · · ,y

) is the objective

vector (a tuple with n objectives). The aim of any

MOGA is to ﬁnd all the solutions for which the

corresponding ﬁtness value can not be improved in

a dimension without degrading another. Therefore,

these algorithms can be applied to set some guidelines

when adjusting the electrodes for neuroestimulation,

reducing the number of trials and saving time and

effort to both pacients and researchers.

4.1 Non-dominated Sorting GA-II

One of the most famous MOGAs is the Non-

dominated Sorting GA-II (NSGA2) (Deb et al.,

2002). This algorithm has been widely used in many

applications providing satisfactory results.

The main characteristic of this algorithm is the

way in which the population of the next generation is

obtained. It performs an efﬁcient sorting of the origi-

nal population and the offsprings generated, obtaining

the different sub Paretos and adding them into the new

population until it is completed.

4.1.1 Encoding a Solution

The ﬁrst step that should be done when applying a GA

is the determination of how an individual encodes a

solution. For the problem tackled in this paper, an in-

dividual consits in a vector of real values where each

gene encondes the three parameters to be optimized.

Figure 1 shows a parameter conﬁguration and how

this is encoded in an individual (Cu=20; PD=60;

IPHI=180; IPI=90; NPulses=84).

Table 1: Representation of a chromosome.

Genotype

Cu PD IPHI IPI NPulses

20 60 180 90 84

4.1.2 Genetic Operators

In this work the standard genetic operators have been

used: Tournament Selection (Miller and Goldberg,

1995), BLX-α (α = 0.3) and Uniform Mutation. For

a real coded genetic algorithm, this type of operators

have shown an adequate performance reaching a bal-

ance between the exploration of the solution space

and the explotation of the good values intervals.

5 EXPERIMENTS

This section presents the results provided by the algo-

rithm for a simulated patient. The model that has to

be optimized was described previously, therefore, the

output of the implemented MOGA is a vector of three

values (Q,Brightness,Onset).

There is a wide research on how these parame-

ters could be tuned, nonetheless, those heuristic can

be always replaced by a manual adjustment based on

a set of experiments. Thus, the given values were the

most adequate after several tests. Due to the non-

deterministic nature of GAs, the algorithm was ex-

ecuted 10 times providing robust results with a few

variation in the results.

In this study the average values of the ten experi-

ments are shown in Table 2 for the objectives consid-

ered: Q, Brightness and onSet.

Table 2: Average values results of the objectives considered

in the algorithm.

Ob jective Value

Q 1.250 nC

Brightness 128 ∈ [0,255]

onSet 64800 µs

As the results showed, the ouput of the algorithm

sets the current in a ﬁxed value, obtaining a wider va-

riety of solutions only for the other two objectives.

PRESELECTION OF NEUROSTIMULATION WAVEFORMS FOR VISUAL PROSTHESES USING GENETIC

ALGORITHMS

193

Optimal Pareto Set

10000

20000

30000

40000

50000

60000

70000

80000

90000

350 550 750 950 1150 1350 1550 1750 1950

onSet

Figure 1: Final Pareto obtained by the algorithm.(Units: on-

Set is expressed in microseconds, and Q is expressed in pC).

6 CONCLUSIONS

The tunning of stimulation parameters for visual neu-

roprostheses still remains as one of the most trou-

blesome issues to be faced due to the large num-

ber of parameters and the wide range of values.

This paper has presented an application of one of

the most famous optimization tools to this problem

such as Genetic Algorithms. These techniques al-

low the researchers to obtain a reduced set of pos-

sible solutions so the range of recommended values

to be tested in vivo is signiﬁcantly reduced. For

the concrete case described in the paper, from the

15*15*15*15*100*255=1,290,937,500 possible so-

lutions, a reduced subset of 16 solutions was obtained.

The implementations were tested over a simulation

software providing satisfactory results, showing how

useful is to apply these techniques in order to improve

the adjustment of the neuroprostheses.

ACKNOWLEDGEMENTS

This work has been partially supported by the

Spanish CICYT Projects TIN2007-60587,TIN2008-

06893-C03-02, and Junta Andalucia Projects P07-

TIC-02768, P06-TIC-02007 and TIC-3928, and Uni-

versity of Jaen Project UJA-08-16-10.

REFERENCES

Casillas, J. and Carse, B. (2009). Special issue on Genetic

Fuzzy Systems: Recent Developments and Future Di-

rections. Soft Computing, 13(5):417–418.

Cord

on, O., Alcal

a, R., Alcal

a-Fdez, J., and Rojas, I.

(2007). Special Issue on Genetic Fuzzy Systems:

What’s Next? Editorial, IEEE Transactions on Fuzzy

Systems, 15(4):533–535.

Deb, K., Pratap, A., Agrawal, S., and Meyarivan, T. (2002).

A fast and elitist multiobjective genetic algorithm:

NSGA-II. IEEE Transactions Evolutionary Compu-

tation, 6(2):182–197.

Dobelle, W. H. (2000). Artiﬁcial Vision for the Blind by

Connecting a Television Camera to the Visual Cortex.

Asaio J, 46:3–9.

Fern

andez, E., Pelayo, F., Romero, S., Bongard, M., Marin,

C., Alfaro, A., and Merabet, L. (2005a). Development

of a cortical visual neuroprosthesis for the blind: the

relevance of neuroplasticity. Journal of Neural Engi-

neering, 2(4).

Fern

andez, E., Pelayo, F., Romero, S., Bongard, M., Marin,

C., Alfaro, A., and Merabet, L. (2005b). Development

of a cortical visual neuroprosthesis for the blind: The

relevance of neuroplasticity. J. of Neural Eng, 4:R1–

R12.

Golberg, D. (1989). Genetic Algorithms in search, opti-

mization and machine learning. Addison-Wesley.

Guillen, A., Pomares, H., Rojas, I., Gonz

alez, J., Valen-

zuela, O., and Prieto, B. (2009). Parallel Multi-

objective Memetic RBFNNs Design and Feature Se-

lection for Function Approximation Problems. Neu-

rocomputing.

Guill

en, A., Rojas, I., Gonz

alez, J., Pomares, H., Her-

rera, L. J., and Fern

andez, F. (2006). Multiobjective

RBFNNs Designer for Function Approximation: An

Application for Mineral Reduction. Lecture Notes in

Computer Science, 4221:511–520.

Humayun, M. S. (2003). Visual perception in a blind sub-

ject with a chronic microelectronic retinal prosthesis.

Vision Res, 43(24):2573–2581.

Isibuchi, H. (2007). Multiobjective genetic fuzzy sys-

tems: review and future research directions. In FUZZ-

IEEE’07, pages 913–918, London.

Miller, B. L. and Goldberg, D. E. (1995). Genetic Al-

gorithms, Tournament Selection, and the Effects of

Noise. Complex System, 9:193–212.

Normann, R. A., Greger, B. A., House, P., Romero, S. F.,

Pelayo, F., and Fernandez, E. (2009). Toward the de-

velopment of a cortically based visual neuroprosthe-

sis. J. Neural Eng., 6(4):049802+.

Romero, S. F., Morillas, C. A., Pelayo, F. J., and Fern

andez,

E. (2008). Computer-controlled neurostimulation for

a visual implant. In BIODEVICES (1), pages 84–91.

Schmidt, E., Bak, M., Hambrecht, F., Kufta, C., O’Rourke,

D., and Vallabhanath, P. (1996). Feasibility of a visual

prosthesis for the blind based on intracortical micros-

timulation of the visual cortex. Brain, 119:507–522.

Troyk, P. e. a. (2003). A Model for Intracortical Visual Pros-

thesis Research. Artif. Organs, 11:1005–1015.

Veraart, C. (1998). Visual sensations produced by optic

nerve stimulation using an implanted self-sizing spi-

ral cuff electrode. Brain Res, 813:181–186.

WHO (2005). World health organization: Prevention of

avoidable blindness and visual impairment.

BIODEVICES 2010 - International Conference on Biomedical Electronics and Devices

194